Iteratarr

AI video iteration and evaluation tool built around the Five Rope methodology. Structured scoring replaces "pretty good" with measurable, traceable, improvable.

What It Does

Every AI video render gets evaluated, scored, attributed to a specific control lever, and linked to its generation settings. Every iteration decision is traceable. Every improvement is measurable. Nothing lives in someone's head.

The core loop:

Generate JSON → Render in Wan2GP → Score in Iteratarr →
Attribute to rope → Generate next iteration JSON → Repeat until locked

The Five Rope Methodology

Six controllable levers (called ropes) drive every prompt engineering decision:

Rope	What it controls	JSON field
1 — Prompt Position	Word order. Identity block first, never compressed.	prompt
2 — Attention Weighting	Token emphasis. (mckdhn:1.3) boosts, (Monaco:0.9) reduces.	prompt
3 — LoRA Multipliers	Phase-aware weights for dual-DiT. "1.0;0.3 0.3;1.2"	loras_multipliers
4a — CFG High Noise	Prompt adherence in composition pass. Sweet spot: 5.9-6.2	guidance_scale
4b — CFG Low Noise	Prompt adherence in identity refinement pass.	guidance2_scale
5 — Steps Skipping	Faster iteration renders. Taylor2 cache. Off for production.	skip_steps_cache_type
6 — Alt Prompt	Secondary prompt driving low noise phase only. Pure identity.	alt_prompt

Bonus levers: flow_shift, NAG_scale, sample_solver

Iteration Protocol

• 32 frames for iteration (half render time)
• 81 frames for production (after lock at 65/75)
• Same seed across all iterations — controlled experiment
• One variable change per iteration — clean comparison
• Score threshold: 65/75 to lock as production ready

Evaluation Scoring

75 points maximum across three categories:

Identity (40 max): Face match, head shape, jaw, cheekbones, eyes/brow, skin texture, hair, frame consistency

Location (20 max): Location correct, lighting, wardrobe, geometry

Motion (15 max): Action executed, smoothness, camera movement

Features

• Three-panel dark UI — terminal meets broadcast suite aesthetic
• Episode Tracker kanban — not started → in progress → evaluating → locked → in queue
• 15-slider evaluation with live score ring and ghost markers (previous iteration scores)
• Smart rope guidance — contextual suggestions based on lowest-scoring element
• Import Evaluation — paste structured JSON from AI assistant, pre-fills everything
• AI vs human score tracking — ai_scores, human_scores, score_deltas on every evaluation
• Side-by-side video comparison — previous vs current render via VideoDiff component
• FFmpeg frame extraction — auto-extract thumbnails, copy path for external review; dedupes PNG/WebP frames and skips corrupted files before vision scoring
• Frame strip — thumbnail strip in evaluation panel via FrameStrip component
• JSON diff panel — see exactly which generation settings changed
• Prompt intelligence — phrase-level prompt diffing, score correlation, and inline diff UI via PromptDiffInline
• Iteration lineage — ancestor chain visualization and scoring pattern analysis via IterationLineage
• Ghost markers — previous iteration scores as trend-coloured dots on sliders
• Character Registry — LoRA files, locked identity blocks, proven settings
• Production lock — 7-step workflow: LOCKED folder, 81-frame JSON, DaVinci sidecar, queue
• Structured file layout — episode/scene/clip hierarchy with auto-naming
• Wan2GP integration — output_filename auto-set, render path auto-detected
• Autopilot — fully automated score → recommend → apply → render loop
• Cross-clip analytics dashboard — overview pills, per-character performance, rope effectiveness bars, seeds analysis tab (SeedsTab), stalling/locked/healthy clip sections; shortcut in Seed HQ header
• Storage monitoring — disk usage tracking per clip, iteration, and frame directory; surfaced in StoragePage
• Telemetry foundation — opt-in, anonymized, off by default

Stack

• Backend: Node.js, Express, SQLite (better-sqlite3), FFmpeg
• Frontend: React 18 (Vite), Tailwind CSS v4
• Ports: 3847 (backend), 3848 (frontend dev)

Quick Start

# Backend
cd backend
npm install
npm run dev

# Frontend (separate terminal)
cd frontend
npm install
npm run dev

Open http://localhost:3848

Configuration

Edit backend/config.json:

{
  "wan2gp_json_dir": "path/to/wan2gp/app",
  "wan2gp_output_dir": "path/to/wan2gp/app/outputs",
  "wan2gp_lora_dir": "path/to/wan2gp/app/loras/wan",
  "project_base_dir": "path/to/your/project",
  "vision_model": "claude-sonnet-4-6",
  "score_lock_threshold": 65,
  "iteration_frame_count": 32,
  "production_frame_count": 81,
  "telemetry_enabled": false,
  "port": 3847
}

Settings Philosophy

Three tiers. Curation is the moat. Not completeness.

Tier 1 — Core Ropes: Always visible. Directly and predictably affect output quality.

Tier 2 — Advanced Levers: Collapsed by default. Power users reach for these when Tier 1 isn't solving the problem.

Tier 3 — Passthrough: In the JSON for Wan2GP. Never shown. Preserved silently on generation.

Telemetry

Off by default. Opt-in only. When enabled, records evaluation scores, rope attributions, and generation settings locally. Export is fully anonymized — no file paths, no prompt text, no character names.

The methodology is universal. The data is model-specific. Keep them cleanly separated.

License

Private. Not yet released.